rune.fast

$17,679 of tokens at list price, on $256 of subscriptions

Five and a half days of an agent swarm, from 2:37 PM on Tuesday 20 October to 4:15 AM on Monday 26 October.

What it cost

The large number is list-price value: what the model providers’ APIs would have charged for the tokens the swarm used. The swarm ran on six Claude Max plans and one ChatGPT Pro plan, $1,400 a month and shared with other projects; these five and a half days are $256 of that. List value is 69 times what was paid.

Thursday cost the most; managers outspent workers on Friday, Sunday and MondayList value per daymanagersworkersthreadscarry and document
$1,114Tue 20$3,803Wed 21$5,448Thu 22$3,490managers aheadFri 23$2,106Sat 24$1,529Sun 25$189Mon 26

Thursday 22 October was the costliest day at $5,448, and workers led it, $2,789 to the managers’ $1,757. Managers are coordinator issues: agents that split a mission into child issues and review what comes back. On Friday they pulled ahead, $1,560 to $956. Saturday went back to the workers, $703 to $546, and Sunday to the managers, $426 to $344.

32 standing issues cost more than the 544 that mergedList value by issue status at the finish
  1. $4,843544 merged issues
  2. $3,137215 closed, landed by snapshot
  3. $5,64632 standing: coordinators and long-lived missions
  4. $3,985six threads

$5.38the median merged issue

Runs tied to no listed issue make up the remaining $68.

Opus carried 72 percent of the list value
ModelList value
Opus$12,779
Fable$3,157
GPT-6 Astra$1,671
GPT-6.1 Sol$37
Haiku$17
Sonnet$14
GPT-5.6 Luna$3
GPT-5.6 Terra$1

Sunday 25 October was the GPT day. GPT-6 Astra took $1,113 of the day’s $1,529, 73 percent, against $238 for Opus.

What the swarm looked like

The swarm was 906 units of work: 900 issues and six long-running threads.

agent runs
3,761
turns
19,967
messages between units
26,898
forced swaps, Claude to GPT and back
36
At the peak, 87 agents produced output in the same five minutesAgents producing output, per five minutes, counted as runs that wrote at least one model response in the bucket, so idle sessions do not count. Shaded and numbered: the five scored attempts.
0204060801 2 345Wed 21Thu 22Fri 23Sat 24Sun 2587 agentsFri 23, 3:07 PM

The peak came at 3:07 PM on Friday 23 October. On Thursday, the costliest day, the busiest five minutes had 61.

Spawning and merging both peaked on Thursday 22 OctoberIssues spawned and merged per day
0100200300269171spawnedmergedTue 20Wed 21Thu 22Fri 23Sat 24Sun 25Mon 26

The swarm spawned 522 issues on 21 and 22 October and merged 341; every later day spawned fewer than either. Messages between units peaked on 22 October too, at 8,840.